Research Information System
6th International Conference on Frontiers in Academic Research, Konya, Turkey, 16 - 17 December 2025, pp.1241-1247, (Full Text)
Condition monitoring of hydraulic systems is a critical component of predictive maintenance strategies, as gradual degradation mechanisms such as internal pump leakage can significantly affect system efficiency and reliability. This study presents a data-driven methodology for classifying internal pump leakage severity using multi-sensor measurements acquired from an experimental hydraulic test rig. The dataset consists of 2205 constant-load operating cycles, each with a duration of 60 seconds, and includes pressure, flow, motor power, temperature, vibration, and efficiency-related signals sampled at multiple rates ranging from 1 Hz to 100 Hz. To address the high dimensionality of the raw data, cycle-level statistical features are extracted and used to train a fully connected deep neural network for three-class classification: no leakage, weak leakage, and severe leakage. Experimental results show that on the training dataset, the model achieves F1-scores of 0.928, 0.655, and 0.816 for no leakage, weak leakage, and severe leakage classes, respectively. On the test dataset, the corresponding F1-scores are 0.897, 0.553, and 0.764, with a severe leakage recall of 0.743. The results demonstrate that the proposed feature-based deep learning framework provides stable generalization performance and reliable detection of advanced leakage conditions, making it suitable as a baseline methodology for hydraulic condition monitoring.